- By Admin
- 08/17/2026 12:15:01
ALL
PMI’s New AI Standard for Project Delivery
From Ad Hoc to Accountable: Applying PMI's New AI Standard Inside Your Project Delivery Stack
There is now an AI project management standard, and it is not a vendor framework. PMI has released The Standard for Artificial Intelligence in Portfolio, Program, and Project Management — the first and only ANSI-approved AI standard written for the project profession. It gives delivery organisations a documented route from ad hoc AI use to structured, accountable practice, covering AI business cases, tool selection, AI-specific risk management, ethics oversight and alignment with obligations such as ISO 42001 and the EU AI Act.
The timing matters because most project-based organisations are already past the point of choosing whether to use AI. Schedulers are drafting programmes with it, bid teams are producing proposals with it, and delivery leads are asking it to summarise status reports. What almost none of them have is a written answer to the questions an auditor, a client or a regulator will ask: which tools are approved, on what data, reviewed by whom, and with what record of the decision.
This article looks at what the standard actually asks of a delivery organisation, why an AI project management standard lands differently across the five markets Arcprojects.io serves, and what a PMO can practically do in the next quarter without stalling the AI work already underway.
Key takeaways
- PMI's standard is the first and only ANSI-approved AI standard written specifically for portfolio, programme and project management.
- It covers five practical domains: AI business cases, tool selection, AI-specific risk management, ethics oversight and regulatory compliance.
- It is written to sit alongside ISO 42001 AI management systems and obligations such as the EU AI Act, rather than duplicating them.
- The evidence any AI governance regime depends on is ordinary delivery data: who logged what time, which budget decisions were made when, which forecasts changed and why.
- AI features in delivery tooling are only as trustworthy as the underlying project records — governance starts with data quality, not model choice.
Why this matters right now across the five target markets
In the United States, the pressure is contractual before it is regulatory. Large enterprise and public-sector buyers have started adding AI-use disclosure clauses to professional services agreements, asking suppliers to declare where AI contributed to a deliverable and to confirm that client data was not used to train third-party models. A supplier without a documented AI tool selection framework answers those questions inconsistently across bids, which is a procurement risk long before it is a compliance one.
In Canada and Australia, the dominant concern among delivery organisations is privacy and data residency in AI tooling, together with professional liability. An engineering consultancy that used a general-purpose model to draft a technical assessment carries the same duty of care it always did, and the standard's insistence on human accountability for AI-assisted outputs maps directly onto professional obligations those firms already understand. In South Africa and India, the practical driver is different again: rapid AI adoption inside delivery teams, often ahead of any organisational policy, in markets where the client base spans domestic and international work with very different expectations.
Across all five, one thing is constant. Firms delivering into the European Union — whether headquartered in Sydney, Toronto, Johannesburg, Bengaluru or Chicago — are within reach of the EU AI Act's extraterritorial scope where their systems or outputs are used there, and ISO 42001 is increasingly appearing in supplier questionnaires as a proxy for AI maturity regardless of jurisdiction. The standard is useful precisely because it gives a project organisation one internal framework that can be evidenced against several external ones, rather than a separate compliance exercise per market.
What it means for PMO leads, delivery directors and quality or risk managers
If you lead a PMO, a delivery function or quality and risk at a project-based firm adopting AI tooling, the standard changes what you are accountable for in three concrete ways.
First, AI adoption needs a business case per use, not a blanket licence decision. Most organisations bought an AI capability at the organisation level and then let usage find its own level. The standard pushes towards a documented case for each significant application — AI-assisted schedule risk analysis, automated status summarisation, resource forecast prediction — with an expected benefit, a defined data input, an owner and a review point. That is not bureaucracy for its own sake; it is what lets you switch off the applications that are producing noise while defending the ones that are producing value.
Second, responsible AI in the PMO requires a named human accountable for every AI-assisted output that leaves the organisation. This is the clause with the sharpest operational edge. A forecast produced with model assistance, a risk register generated from historical data, a client status report drafted automatically — each needs an identified reviewer who accepted it. Organisations that cannot show who approved a given output cannot claim human oversight, whatever their policy document says.
Third, AI-specific risk has to enter the risk register as its own category. Model drift, hallucinated content in client deliverables, data leakage through prompts, over-reliance on automated forecasting and vendor concentration are risks with different treatment paths from the delivery risks a PMO usually manages. The standard's contribution here is mostly taxonomic, and that is genuinely useful: it stops AI risk being filed as a vague technology line item that nobody owns.
Three practical implications for your delivery model
Governance runs on delivery records
Every control the standard describes needs evidence — approvals, timestamps, versioned forecasts. A PMO without an auditable project record cannot demonstrate oversight regardless of the policy it has written.
Tool selection becomes a documented decision
Which tools are approved, for what data classification, with what retention terms and which fallback if the vendor changes model. Written once, applied per use case, reviewed on a cycle.
Forecast quality is a data problem first
AI-assisted overrun prediction and capacity forecasting depend on complete hours, clean budget baselines and honest progress reporting. Poor inputs make a confident model worse than no model.
Spreadsheets, point tools or a connected delivery platform
| Capability | Spreadsheets | Point tools | Arcprojects.io |
|---|---|---|---|
| Auditable approval trail | Version-dependent, easily overwritten | Partial, per tool | Approver, timestamp and comment retained across modules |
| Versioned forecast history | Overwritten each update | Rarely retained | Baseline and revision history preserved |
| Single source of delivery truth | Scattered across files and owners | Fragmented across three or four systems | Schedule, cost, effort and capacity in one record set |
| Data quality for AI-assisted forecasting | Inconsistent, manually keyed | Good in one domain, absent elsewhere | Structured, complete and continuously updated |
| Portfolio-level oversight | Manual consolidation | Project-level only | Dashboards across every live programme |
| Evidence for a client or audit request | Days of reconstruction | Export and merge from several tools | Reported on demand from one source |
| Change and decision traceability | Buried in email threads | Ticket history, disconnected from cost | Decisions recorded against schedule and budget |
How Arcprojects.io helps you evidence AI governance in project delivery
Arcprojects.io provides the control evidence layer any AI standard assumes exists: Dashboards and Reporting for portfolio oversight, Gantt Charts for baselined and version-tracked schedules, and Cost Control for budget decisions recorded with an owner and a date. The standard does not ask you to buy an AI product; it asks you to be able to show what happened, who decided it and on what basis. That is a project records problem, and it is exactly what a connected delivery platform produces as a by-product of normal work.
- Establish one auditable delivery recordBring schedule, cost, effort and approvals into a single system so that every forecast change, budget revision and time entry carries an owner and a timestamp. This is the substrate for human-oversight evidence: when someone asks who accepted an AI-assisted output, the answer is in the approval history rather than in someone's memory.
- Baseline and version your forecastsUse Gantt baselines and cost baselines so that revisions are recorded rather than overwritten. A governance review needs to see what the forecast said before and after an AI-assisted adjustment, and an organisation that overwrites its plans cannot demonstrate that any oversight occurred at all.
- Report oversight at portfolio levelUse dashboards to review AI-assisted decisions across the portfolio on a fixed cycle — which programmes used which applications, what changed as a result, and where variance moved afterwards. That review, minuted and repeated, is what turns a written AI policy into demonstrable practice.
See what auditable delivery evidence looks like
Walk through baselined schedules, cost decision trails, approval history and portfolio dashboards with a specialist who works with project-based firms.
Request a demo"An AI standard does not ask whether your model is good. It asks whether you can show who was accountable for what it produced — and that is a records question, not a technology question."
"We had an AI policy and no evidence. The first client questionnaire asked which tools we used on their programme and who reviewed the outputs, and we spent a fortnight assembling an answer. Now the approval trail already exists in the delivery record, so the questionnaire takes an afternoon instead."
PMO lead's AI governance readiness checklist
- List every AI tool currently in use across delivery teams, including the ones nobody formally approved.
- Classify the data each tool may process, and write down what must never be entered into a third-party model.
- Write a business case per significant AI application, with an owner, an expected benefit and a review date.
- Name a human accountable for reviewing every AI-assisted output that reaches a client.
- Add AI-specific risks — drift, hallucination, data leakage, over-reliance, vendor concentration — to the risk register as their own category.
- Baseline and version schedules and budgets so forecast changes are recorded, not overwritten.
- Check whether any of your delivery reaches the European Union and assess EU AI Act exposure accordingly.
- Map your internal controls once against ISO 42001 so supplier questionnaires can be answered from a single source.
- Set a fixed portfolio-level review cycle for AI-assisted decisions, and minute it.
Frequently asked questions
What is PMI's standard for artificial intelligence in project management?
It is The Standard for Artificial Intelligence in Portfolio, Program, and Project Management, the first and only ANSI-approved AI standard written for the project profession. Rather than prescribing tools, it describes the practices a delivery organisation should have in place to move from ad hoc AI use to accountable practice: how to build an AI business case, how to select and review tools, how to identify and treat AI-specific risks, how to maintain ethics and human oversight, and how those practices align with wider obligations such as ISO 42001 and the EU AI Act. Full details are in the PMI announcement of the standard.
How do you govern AI use across a project portfolio?
Start with an inventory rather than a policy. Establish which AI applications are actually in use across programmes, what data each touches, and what outputs reach clients. Then apply three controls: an approved tool list with data classifications, a named human reviewer for any AI-assisted output that leaves the organisation, and a portfolio-level review on a fixed cycle that examines what changed as a result of AI-assisted decisions. The common failure is writing the policy first and discovering later that half the delivery teams were using tools it never contemplated. Governance that is not grounded in an accurate picture of current practice tends to be ignored within a quarter.
What does the EU AI Act mean for project delivery teams?
Its relevance depends on where your systems and outputs are used rather than where your firm is registered, so delivery organisations in any of the markets Arcprojects.io serves may fall within scope when they deliver into the European Union. In practice, project teams should expect obligations around transparency of AI use, risk classification of the applications they deploy, documentation of how systems were selected and tested, and demonstrable human oversight of outputs. The PMI standard is deliberately written to sit alongside these obligations rather than to replace them, which is why mapping your internal controls once is more efficient than responding to each framework separately. Background on how the standard addresses regulatory alignment is set out in the PMI release on the global AI standard.
Accountability is built from ordinary records
The organisations that will find this standard straightforward are not the ones with the most advanced AI capability. They are the ones whose delivery records already answer basic questions: what the plan said last month, who changed it, what the budget assumed, and who approved the hours behind the cost. AI governance sits on top of that. Where it does not exist, an AI policy is a statement of intent that cannot be evidenced, and the first serious client questionnaire will make that obvious.
Arcprojects.io keeps schedule, cost, effort and approvals in one place, which is what makes an oversight claim demonstrable rather than assertive. Explore how the dashboard, reporting, Gantt and cost control modules work together on the features and benefits page, see the kinds of organisations already running on it at who uses Arcprojects, or review pricing. You can also start a 30-day free trial and test whether one live programme can produce its own audit trail on demand.